---
license: mit
datasets:
- FronkonGames/steam-games-dataset
metrics:
- accuracy
base_model:
- google/efficientnet-b3
pipeline_tag: image-classification
tags:
- game
---
# 🎮 GameNet-1

**GameNet-1** is a deep learning-based computer vision system designed to recognize video games based on their cover art or in-game screenshots. Built using EfficientNet and trained on a curated dataset of popular Steam games, the model predicts both the **game name** and its **genre(s)**.

---

## 🚀 Features

- 🔍 Recognizes games from screenshots or cover images
- 🧠 Powered by EfficientNetB3 for high accuracy
- 🗂️ Trained only on **popular games** with over 2M estimated owners
- 🎯 Fine-tuned and augmented for better generalization
- 📊 Shows prediction confidence alongside game metadata

---

## 📁 Dataset

- Source: [Steam Games Dataset on Kaggle](https://www.kaggle.com/datasets/fronkongames/steam-games-dataset)
- Filtered for popular games with over 2 million estimated owners
- Images:
  - Header cover image
  - 5 in-game screenshots (JPEG only)

---

## 🏗️ Model Architecture

- **Base**: `EfficientNetB3` pretrained on ImageNet
- **Input Size**: 300x300 RGB
- **Top Layers**:
  - `GlobalAveragePooling2D`
  - `Dropout` (0.4 & 0.2)
  - `Dense(256, relu)`
  - `Dense(n_classes, softmax)`
- **Training**:
  - Phase 1: Frozen base
  - Phase 2: Fine-tuned base (lower LR)

---

## 📈 Performance

- Accuracy (val set): 30%
- Trained using:
  - `categorical_crossentropy` loss
  - `Adam` optimizer (1e-3 for frozen, 1e-5 for fine-tune)
  - Real-time data augmentation (`ImageDataGenerator`)

---